Model Card: assets overview, quick start, license, citation

README.md

@@ -1,3 +1,86 @@
 ---
 license: cc-by-nc-4.0
+language:
+- en
+tags:
+- human-trajectory-prediction
+- social-transmotion
+- jrdb
+- jta
+- cvpr2025
+- emloco
+pipeline_tag: other
+library_name: pytorch
 ---
+
+# EmLoco — CVPR 2025 release assets
+
+This repository hosts the **`Ours` (num_modes=1) checkpoints** and **CVPR-era preprocessed shards** that accompany:
+
+> **Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment** (CVPR 2025).
+> Hiromu Taketsugu, Takeru Oba, Takahiro Maeda, Shohei Nobuhara, Norimichi Ukita.
+> [📄 paper](https://openaccess.thecvf.com/content/CVPR2025/html/Taketsugu_Physical_Plausibility-aware_Trajectory_Prediction_via_Locomotion_Embodiment_CVPR_2025_paper.html) · [arXiv 2503.17267](https://arxiv.org/abs/2503.17267) · [project page](https://iminthemiddle.github.io/EmLoco-Page/) · [🐙 source code](https://github.com/ImIntheMiddle/EmLoco)
+
+## What's here
+
+| Path | Size | Content |
+|---|---|---|
+| `checkpoints/jta_ours/` | 38 MB | JTA Ours model (num_modes=1, EmLoco loss, token_num=49) — `best_val_checkpoint` of `jta_valuenet_100` |
+| `checkpoints/jrdb_ours/` | 38 MB | JRDB Ours model (num_modes=1, EmLoco loss, token_num=26) — last `checkpoint` of `jrdb_value150` |
+| `preprocess_smpl_cvpr/jta_all_visual_cues/{train,val,test}/` | 23 GB | JTA J=49 shards (torch 2.x zip-format `.pt`, 21 parts) |
+| `preprocess_smpl_cvpr/jrdb_all_visual_cues/{train,val,test}/` | 4.5 GB | JRDB J=26 shards (`.pkl`, 7 parts). Pose tokens are NaN-filled for frames without a JRDB-Act label so the EmLoco loss skips them automatically. |
+
+Each checkpoint directory also contains a `config.yaml` capturing the training-time hyper-parameters.
+
+## Quick start
+
+```bash
+# 0. Clone EmLoco source + set up the unified Python 3.8 / CUDA 12.1 env
+git clone https://github.com/ImIntheMiddle/EmLoco && cd EmLoco
+uv sync && source .venv-22.04/bin/activate
+
+# 1. Pull the assets from this repo
+pip install -U "huggingface_hub[cli]"
+hf download iminthemiddle/EmLoco --local-dir .assets --repo-type model
+
+# 2. Wire them into the expected paths
+ln -s "$PWD/.assets/preprocess_smpl_cvpr/jta_all_visual_cues" \
+      social-transmotion/data/jta_all_visual_cues/preprocess_smpl_cvpr
+ln -s "$PWD/.assets/preprocess_smpl_cvpr/jrdb_all_visual_cues" \
+      social-transmotion/data/jrdb_all_visual_cues/preprocess_smpl_cvpr
+mkdir -p social-transmotion/experiments/{JTA,JRDB}
+ln -s "$PWD/.assets/checkpoints/jta_ours"  social-transmotion/experiments/JTA/jta_ours
+ln -s "$PWD/.assets/checkpoints/jrdb_ours" social-transmotion/experiments/JRDB/jrdb_ours
+
+# 3. Grab the LocoVal value-network from the GitHub Release (28 KB)
+mkdir -p pacer/output/exp/pacer
+# https://github.com/ImIntheMiddle/EmLoco/releases/tag/checkpoints  (unzip valuenet_checkpoints.zip)
+
+# 4. Evaluate the released Ours checkpoints
+python social-transmotion/evaluate_jta.py  --exp_name jta_ours  --modality traj+all   # ADE 0.951 / FDE 1.921
+python social-transmotion/evaluate_jrdb.py --exp_name jrdb_ours --modality traj+all   # ADE 0.369 / FDE 0.724
+```
+
+## Reproduced numbers (JRDB-Traj test, JTA-Dataset test)
+
+| Setting | ADE | FDE |
+|---|---|---|
+| JTA Ours (`jta_ours`, num_modes=1) | **0.951** | **1.921** |
+| JRDB Ours (`jrdb_ours`, num_modes=1) | **0.369** | **0.724** |
+
+For the multi-modal Table rows (num_modes>1, LocoVal filter at inference), train your own with `--multi_modal --valueloss_w 1.0` on top of `preprocess_smpl_cvpr/` or contact the authors for additional checkpoints.
+
+## License
+
+The assets in this repository are released under **CC BY-NC 4.0** — *research, non-commercial use only*. The reason is that the underlying SMPL body model and the JTA / JRDB datasets used during preprocessing and training are themselves restricted to non-commercial research use, as is the PACER (NVIDIA) code that produced the LocoVal value function. The EmLoco source code on GitHub remains under MIT.
+
+## Citation
+
+```bibtex
+@InProceedings{EmLoco_CVPR25,
+  author    = {Taketsugu, Hiromu and Oba, Takeru and Maeda, Takahiro and Nobuhara, Shohei and Ukita, Norimichi},
+  title     = {Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment},
+  booktitle = {IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR)},
+  year      = {2025}
+}
+```